1.1 Key Contributions

This work directly addresses the problem of overreliance on labeled datasets in AI, enabling scalable and cost-effective learning in data-scarce domains. By leveraging vast amounts of available unlabeled data, the proposed SSL approach paves the way for practical AI applications in fields such as healthcare, environmental monitoring, and scientific discovery. Future research will focus on extending this approach to multi-modal learning and real-world deployment in resource-constrained environments.

2.1 Self-Supervised Learning Techniques

Self-supervised learning (SSL) has emerged as a powerful paradigm for learning representations from unlabeled data. Early SSL methods focused on pretext tasks, such as predicting image rotations [12] or solving jigsaw puzzles [13], to generate pseudo-labels for training. More recent approaches, such as contrastive learning and masked autoencoding, have demonstrated significant improvements in representation quality.

Contrastive learning techniques, such as SimCLR [2] and MoCo [14], train models by maximizing agreement between differently augmented views of the same image while ensuring separation from negative samples. These methods have been widely used in natural image processing but require large batch sizes or memory banks to be effective.

Masked autoencoders (MAE) [6] have recently gained traction in SSL for vision tasks. They operate by randomly masking portions of the input and training the model to reconstruct the missing regions, thereby capturing meaningful semantic information. Despite their success, contrastive learning and masked autoencoding have rarely been combined for representation learning in data-scarce environments, particularly in remote sensing applications.

2.2 SSL in Remote Sensing and Satellite Imagery

SSL has also been explored in remote sensing applications, where labeled data is often scarce. Unlike natural images, satellite imagery consists of multiple spectral bands with varying spatial resolutions, making SSL more challenging. Researchers have adapted SSL techniques for land cover classification, object detection, and change detection in remote sensing.

For example, Montanaro et al. [15] proposed a semi-supervised learning approach for land cover mapping using limited labeled data, while Stojnic and Risojevic et al. [16] applied contrastive learning to satellite imagery to improve feature extraction. However, many existing SSL methods rely on domain-specific augmentations, such as spectral transformations or geometric manipulations, that may not generalize well to other datasets or tasks.

The EuroSAT dataset [11] has been widely used as a benchmark for SSL in remote sensing. Most prior approaches rely on substantial labeled data for fine-tuning, limiting their real-world applicability. In contrast, the proposed framework focuses on learning robust representations from extremely limited labeled data while leveraging large amounts of unlabeled satellite imagery.

2.3 Leveraging Limited Labeled Data

Beyond SSL, several strategies have been proposed to address the challenge of limited labeled data, including semi-supervised learning, transfer learning, and active learning.

Semi-supervised learning methods, such as MixMatch [17] and FixMatch [18], combine labeled and unlabeled data using consistency regularization and pseudo-labeling. These techniques have shown promising results but still require a significant amount of labeled data for effective performance.

Transfer learning, where pre-trained models are fine-tuned on smaller labeled datasets [19], is another common approach. While transfer learning can improve performance in data-scarce scenarios, it requires extensive labeled datasets for pretraining, which may not always be available for remote sensing.

Active learning [20] is another strategy that reduces annotation costs by selecting the most informative samples for labeling. However, it still requires human intervention, making it less scalable for large datasets.

2.4 Gaps and Contributions

While existing SSL methods have demonstrated impressive results, they often assume access to a moderate amount of labeled data for fine-tuning or evaluation. In contrast, this work focuses on scenarios with extremely limited labeled data, where traditional SSL and semi-supervised methods may struggle.

The proposed approach contributions address these gaps by:

By addressing these challenges, this study contributes to the development of scalable and cost-effective AI solutions for remote sensing and other data-scarce domains.

3.1 SSL Pre-Training Phase

The pre-training phase aims to learn meaningful representations from unlabeled data using a combination of contrastive learning and masked autoencoding. Given an unlabeled dataset ${𝒟}_{\text{unlabeled}}={\left\{x_i\right\}}_{i=1}^N$, the proposed SSL framework applies the following learning strategies.

3.2 Fine-Tuning Phase

After pre-training, the learned representations are fine-tuned using a limited labeled dataset ${𝒟}_{\text{labeled}}={\left\{(x_i,y_i)\right\}}_{i=1}^M$. The encoder $f(\cdot )$ is initialized with pre-trained weights, and a classification head $h(\cdot )$ is added for label prediction.

The model is trained using cross-entropy loss:

3.3 Computational Complexity Analysis

The computational complexity of the proposed SSL framework is analyzed to assess scalability in resource-constrained environments. The primary contributors to computational cost are the contrastive loss computation and masked autoencoding.

Contrastive Learning Complexity: The contrastive loss requires computing similarity scores between all pairs in a batch, leading to a complexity of $O\left(B^2\right)$, where $B$ is the batch size. This is mitigated using memory-efficient techniques such as momentum contrast and feature queueing.

Masked Autoencoding Complexity: The masked autoencoding process involves partial forward passes and reconstruction, which scales as $O\left(B\cdot P\right)$, where $P$ is the percentage of masked pixels. By limiting the mask ratio to 50%, this study maintain computational efficiency.

Overall Complexity: Given a dataset with $N$ images and an encoder-decoder architecture, the total complexity is:

which remains manageable for large-scale training on modern GPUs.

3.4 Evaluation Metrics

The performance of the proposed proposed SSL framework is evaluated using the following metrics:

3.5 Implementation Details

The proposed framework is implemented using PyTorch. The encoder $f(\cdot )$ is based on a ResNet-50 architecture, while the decoder $g(\cdot )$ consists of transposed convolutional layers for upsampling. The model is trained using the Adam optimizer with a learning rate of $1\times {10}^{-4}$ and a batch size of 64. Hyperparameters were tuned using a small validation set.

3.6 Algorithms

Below are the pseudocode algorithms for SSL pre-training and fine-tuning.

To provide a clearer understanding of the proposed SSL framework, this study present a schematic diagram in Figure 1. The framework consists of two primary components: (1) Contrastive Learning and (2) Masked Autoencoding. Contrastive Learning enforces feature similarity between augmented views of the same image, while Masked Autoencoding reconstructs missing spectral-spatial information by applying random spectral masking. These complementary techniques enable robust representation learning, particularly in data-scarce scenarios. The pre-trained encoder is subsequently fine-tuned for downstream land cover classification, demonstrating the effectiveness of self-supervised learning in remote sensing applications.

Figure 1 Proposed Self-Supervised Learning (SSL) Framework. The model leverages contrastive learning to learn robust feature representations while masked autoencoding ensures spatial-spectral consistency. The final fine-tuned model is used for downstream land cover classification.

4.1 Dataset

This study evaluate the proposed framework on the EuroSAT dataset [11], which consists of 27,000 labeled and geo-referenced Sentinel-2 satellite images spanning ten land cover classes: Industrial, Residential, River, Highway, Pasture, Forest, Annual Crop, Permanent Crop, Herbaceous Vegetation, and Sea Lake. The images have a resolution of $64\times 64$ pixels and contain 13 spectral bands, making them a suitable benchmark for self-supervised representation learning.

Dataset Splitting: The dataset is split into:

4.2 Experimental Setup

The proposed SSL framework is implemented using PyTorch and trained on an NVIDIA A100 GPU with 40GB memory. The backbone encoder $f(\cdot )$ is a ResNet-50, while the decoder $g(\cdot )$ consists of transposed convolutional layers for masked autoencoding.

Training Details:

4.3 Baselines

To compare the proposed method with existing approaches, this study evaluate the following baselines:

4.4 Evaluation Metrics

To comprehensively evaluate model performance, this study use the following metrics:

4.5 Results and Analysis

The performance of the proposed framework under different labeled data regimes (1%, 5%, 10%, and 100%) is shown in Table 1.

The proposed SSL framework achieves superior performance across all labeled data settings. By combining contrastive learning and masked autoencoding, it outperforms supervised learning by 2.7% and MixMatch by 2.1%. The enhanced mIoU scores indicate that the model effectively learns more robust feature representations, leading to improved land cover classification accuracy.

4.6 Ablation Study

To examine the contribution of individual components, an ablation study is conducted, as shown in Table 2.

Both contrastive learning and masked autoencoding significantly contribute to the model’s performance. Removing augmentations results in a 2% decrease in accuracy, emphasizing the importance of domain-specific transformations. Ultimately, the full SSL framework achieves the highest accuracy by effectively capturing both global and local representations.

4.7 Computational Complexity Comparison

Table 3 presents a comparison of training time per epoch and inference latency across different methods.

Observations:

4.8 Visualization of Learned Representations

To further illustrate the effectiveness of the proposed SSL framework, this study provide t-SNE visualizations of learned feature representations in Figure 2. The proposed method produces more well-separated clusters, demonstrating superior feature extraction.

Figure 2 t-SNE visualization of the learned representations under different labeled data regimes.

5.1 Interpretation of Results

The results indicate that the proposed SSL framework significantly improves performance in data-scarce scenarios. With only 10% of the labeled data, the proposed framework achieves 81.2% accuracy, outperforming supervised learning by 2.7% and semi-supervised methods by 2.1%. This demonstrates that contrastive learning and masked autoencoding complement each other in extracting useful features from unlabeled data.

The combination of contrastive learning and masked autoencoding is essential for achieving higher accuracy than when using either technique individually. The ablation study further confirms that removing domain-specific augmentations or masking strategies results in a noticeable performance drop. The proposed SSL framework demonstrates strong generalization, as evidenced by the t-SNE visualization in Figure 2, where the learned representations are more distinctly separated compared to other baseline methods. Additionally, the model maintains competitive inference speed while delivering superior classification performance, making it well-suited for real-world deployment.

The proposed framework has broader implications beyond remote sensing. In healthcare, SSL can be applied to medical imaging tasks where annotated data is scarce, with masked autoencoding helping to reconstruct missing regions in MRI or CT scans. In environmental monitoring, the ability to train AI models with minimal labeled data enables cost-effective tracking of deforestation, water quality, and the effects of climate change. Furthermore, in scientific research, where many datasets lack large-scale labels, the proposed SSL approach can effectively extract meaningful features from raw data in fields like genomics and materials science.

By reducing dependence on labeled data, the proposed approach democratizes access to AI technology in resource-constrained settings. Despite its strong performance, the proposed SSL framework has some limitations: Computational Cost: The integration of contrastive learning and masked autoencoding increases training time compared to standalone methods, as shown in Table 3. However, inference speed remains competitive. Hyperparameter Sensitivity: The effectiveness of contrastive learning depends on the temperature parameter $\tau$ and augmentation choices. Masked autoencoding also requires tuning the masking ratio. Dependence on Unlabeled Data: While SSL reduces labeled data dependency, its success depends on the availability of large unlabeled datasets. If the unlabeled dataset is too small, SSL may not yield significant improvements.

Several directions for future research could enhance the effectiveness of the proposed SSL framework: Multi-Modal Learning: Future work can explore integrating additional data sources, such as combining satellite imagery with geospatial metadata, LiDAR data, or weather patterns. Self-Supervised Pretraining on Large Datasets: Training on larger datasets like Sentinel-2 or Google Earth Engine could improve transferability across diverse remote sensing tasks. Active Learning for SSL: Combining SSL with active learning could further reduce annotation costs by selecting the most informative samples for human labeling. Theoretical Analysis: A more formal theoretical investigation into why the combination of contrastive learning and masked autoencoding is effective could provide deeper insights into SSL for remote sensing. Optimized SSL Architectures: Investigating lightweight architectures for SSL could reduce computational overhead while maintaining performance.

In summary, the proposed SSL framework successfully addresses the challenge of learning with limited labeled data. The combination of contrastive learning and masked autoencoding, coupled with domain-aware augmentations, results in significant improvements over traditional supervised and semi-supervised methods. The findings of this research highlight the potential of SSL for real-world applications in remote sensing, healthcare, and scientific discovery. Future research will focus on further optimizing the framework, expanding its applicability to multi-modal learning, and exploring its impact in various domains.